OMX innovation · AI Pathway · v0.1

Augment the value stream.
Not the seat licence.

The default AI playbook is "buy everyone a Copilot licence and hope productivity lifts." That's a procurement decision, not a transformation. OMX takes the opposite path — start with the value stream, find the friction and noise, then design a small, repeatable, governed agent that removes it. Same data. Same people. Less drag.

Walkable 3-page deck: ../ai-augmented-org/  ·  Back to OMX Mid Year Updates library  ·  Live at pitchdecks.plex.nz

1 · The choice

Anti-pattern
Seat-licence sprawl

Every staff member gets a Copilot / Claude / chat tool. Adoption is uneven. Output quality is uneven. Governance is unclear. Sensitive data leaks via opportunistic copy-paste. The bill compounds; the value doesn't.

The OMX pattern
Value-stream augmentation

Map the value stream. Pin the highest-friction step. Build a governed agent for that step — role-aware, business-aware, live-data-fed. Roll out the process, not the licence. Measure the friction removed, not the seats sold.

2 · The model

Five moves, in order. Skip one and the system falls back to seat-licence sprawl.

01 IDENTIFY
Map the value stream
Pick one stream end-to-end. Quote -> cart -> invoice. Lead -> opp -> close.
02 LOCATE
Pin the friction
Where does work wait? Where is data re-entered? Where does someone email a screenshot?
03 ENCODE
Repeatable process
Same inputs, same outputs, every time. If it's not repeatable yet, it's not ready for an agent.
04 AUGMENT
Governed agent
Role context + business knowledge + live data. Auditable, reversible, scoped to one job.
05 MEASURE
Friction removed
Cycle time before vs after. Error rate. Cost per transaction. The stream's own KPIs, not licence counts.

What every agent must carry

Secure
Identity bound to the person and role. Tenant-scoped. No shared service accounts. PII on the same rails as everything else.
Governed
Versioned prompts, change control, audit logs. Every output traceable to inputs. Approval gates where action is material.
Controlled
Bounded scope — does one job. Cannot do other people's jobs. Cannot escalate. Fails closed.
Locked CLAUDE.md
Every agent runs against a managed, version-pinned CLAUDE.md — rules, hard constraints, business glossary. The agent's behaviour is config, not improvisation.
CrowdStrike on endpoint
Same endpoint-protection layer the rest of OMX runs on. Agent activity is visible to the SOC. No off-fleet machines, no shadow installs.
Role context
Knows who is asking, what they're allowed to see, what their team owns, what the next step looks like for them.
Business understanding
OMX taxonomy, sector codes, customer hierarchy, replen rules, DC layout, contract floors. Institutional knowledge encoded.
Live data feeds
Reads Snowflake / Pronto / Salesforce in real time. Not a copy. Not a snapshot. Same numbers the human would see.

3 · Reference architecture

How the requests actually flow. Front of house on the left, the routing brain in the middle, the specialist sub-agents and tools on the right.

Entry points Routing brain Sub-agents MuleSoft MCP Tools & data Microsoft 365 Teams · Outlook · Copilot OMX web apps Lens · Accelerate · WhoOnSite Salesforce Service Cloud · CRM Pronto ERP · order desk Amazon Connect Voice · IVR · agent assist officemax.co.nz Customer + Ask Max Core Routing Agent Identity + role + policy Intent classification Sub-agent selection Output policy + audit Single front door Document / RAG agent PDFs · contracts · SOPs QBR / workflow agent Briefing Room · 183 dives Search / catalog agent Web search · PDP · onboarding Quoting / pricing agent BTS · contract floors DC ops agent Replen · labour · twin Comms / social agent Lifecycle · GenAI social General assistant Fallback · everyday tasks MuleSoft MCP bus Signed Observed Rate-limited Every tool call goes through here Bedrock AWS model gateway Anthropic API Claude model family Snowflake Single source of truth dbt Modelled facts + marts n8n Workflow orchestration Postgres App state Pronto · SF · M365 System-of-record APIs External MCP servers Vendor catalogs · search Governance + security spine — inherited by every agent before it runs a single instruction Locked CLAUDE.md Per-agent rules + business glossary Version-pinned · change-controlled CrowdStrike on endpoint Same SOC-visible endpoint protection No off-fleet machines, no shadow installs Entra identity + role Tenant-scoped, role-aware No shared service accounts Audit log + change control Every output traceable to inputs Approval gates on material actions
User entry → Core Routing Agent → specialist sub-agent → MuleSoft MCP bus → the right tool (Bedrock · Anthropic · Snowflake · dbt · n8n · Postgres · system-of-record APIs · external MCP)
click to expand →
Reference architecture — full size

4 · Where we start — eight OMX value streams

Not theory. Real streams already in flight or queued where the pattern replaces friction we can name. Each is a candidate Phase 1 build.

01 Customer service
ISAM call handling
Friction: no CRM context on inbound, manual lookup.
Augment: Amazon Connect screen-pop with SF summary + agent assist. Live in Accelerate.
02 Quote turnaround
Self-service quoting (BTS)
Friction: small customers ring sales for items already catalogued. ~1 FTE reclaimed.
Augment: guided quoting agent reading contract pricing, MOQs, stock, lead time live.
03 Account research
Briefing Room · QBR pre-fill
Friction: half a day of manual assembly per QBR. Data exists; pre-fill is manual.
Augment: Company Research agent reading Snowflake + Salesforce + 183 deep dives.
04 Product onboarding
New SKU enrichment
Friction: bare-minimum attributes land; search misses; PDP looks thin; off-range margin leaks.
Augment: enrichment agent reading vendor spec + Icecat + competitor PDPs; classifies to L5; flags low-confidence for review.
05 Digital comms · social
Lifecycle + GenAI social
Friction: campaigns one-off, brand voice drifts, social cadence depends on who has time.
Augment: comms agent with brand tokens + audience segments live from Snowflake; drafts to a stewarded human queue.
06 Internal SaaS build
Lens · Accelerate · WhoOnSite · Surface
Friction: packaged SaaS for every internal need = recurring licence + vendor captivity.
Augment: build agents (proven on Lens, Accelerate, WhoOnSite) — Snowflake-native, OMX design language, weeks not quarters.
07 Range planning
Product + category reviews
Friction: cadence is calendar-driven, not signal-driven; decisions lag the data.
Augment: review agent with live sell-through, margin, stock-cover, OTIF, competitor coverage; pre-builds the review pack.
08 Switching · matching
"Match my old supplier" + sustainability
Friction: manual product-equivalence matching today. ~2 FTE reclaimed combined.
Augment: matching agent finds OMX equivalents from competitor SKU lists; sustainability scoring; switches the cart.

5 · The AI SDLC — what we've already built

The Pathway isn't theory. The same pattern that augments the value stream also augments our software lifecycle — Claude in the loop on every spec, every build, every review. The platforms below were all delivered through this AI-augmented SDLC: small focused agents, locked CLAUDE.md per project, governed by the same spine on the diagram above.

Lens
OMX BI platform on Snowflake
In production. Replaces Sisense ($130k/yr reclaim from January). Same data, better UX. The classifier-grade proof of fit-for-OMX over packaged SaaS.
Accelerate
EPMO portfolio app — the PMO is the demo
Live PMO at accelerate.officemax.co.nz. Tier-aware governance, RAG, RASCI, milestones, status. The PMO process is now the codebase. ~$30k/yr packaged-PM reclaim.
WhoOnSite
NFC + Gallagher frontline identity
In production. AD-only workers without SF licences finally have a system-of-presence. Input to DC Optimisation, Mobile Workforce, Desk Booking.
Ribbons (IT Help Desk Discovery)
File-move mechanic + operational sequence
AI-led discovery of the IT Help Desk intake process; ribbons design, operational sequencing, technical lead clarified. Feeds the AWS Connect IT Help Desk migration.
Stage 1 Diagnose
Servicing the Unmanaged — programme diagnosis
AI-assisted diagnosis of the unmanaged-540k opportunity. The base layer underneath the six-stage roadmap (Diagnose → Web Opt → Switching → Reorder → Marketplace → The Ball).
Briefing Room + 183 Deep Dives
Company Research engine
Front-end + QBR process + 183 strategy dives on big NZ companies, all built in the AI SDLC. The augmented account-research stream proven in code.
Plex-CI scraper
Competitor intelligence sitemap engine
Sitemap channel unblocked 2026-06-27: 545,789 URLs and 471,991 change events across 59 competitors. Built in days, not months, through AI-paired implementation.
JARVIS + 181-agent harness
The internal AI control plane
3-layer agent architecture (Router → 14 Supervisors → 163 Specialists). The framework that orchestrates the SDLC we're describing. Eating our own cooking.
The point: none of this required a transformation programme to start. Each platform was a small focused build by a small focused team with AI in the loop — same pattern as the value-stream agents we're proposing for the next wave.

6 · The detailed stories

Each link below opens the deep narrative for a programme already in the AI Pathway shape — value stream identified, friction named, governed agent designed or shipping. Walk any one to see the pattern in real OMX context.

7 · The article

Sir — which article do you want linked here? Candidates I can wire in right now:

JARVIS · built 2026-06-29 for the Stella visit · Lens design language · no mocks, no invented numbers